Original contribution
Spatial normalization of multiple sclerosis brain MRI data depends on analysis method and software package

https://doi.org/10.1016/j.mri.2020.01.016Get rights and content

Highlights

  • Compared various spatial normalization approaches using brain MRIs from MS participants and a ‘simulated lesion’ dataset

  • Nonlinear warping systematically outperformed conventional linear (affine) normalization, with SPM (CAT12) providing the most consistent results among evaluated methods

  • Lesion-filling improved spatial normalization accuracy for each method, but these effects were relatively small compared to differences between normalization algorithms

  • SPM (CAT12) with lesion-filling was found to be the most robust approach for spatially normalizing MRI data with MS pathologies

Abstract

Background

Spatially normalizing brain MRI data to a template is commonly performed to facilitate comparisons between individuals or groups. However, the presence of multiple sclerosis (MS) lesions and other MS-related brain pathologies may compromise the performance of automated spatial normalization procedures. We therefore aimed to systematically compare five commonly used spatial normalization methods for brain MRI – including linear (affine), and nonlinear MRIStudio (LDDMM), FSL (FNIRT), ANTs (SyN), and SPM (CAT12) algorithms – to evaluate their performance in the presence of MS-related pathologies.

Methods

3 Tesla MRI images (T1-weighted and T2-FLAIR) were obtained for 20 participants with MS from an ongoing cohort study (used to assess a real dataset) and 1 healthy control participant (used to create a simulated lesion dataset). Both raw and lesion-filled versions of each participant's T1-weighted brain images were warped to the Montreal Neurological Institute (MNI) template using all five normalization approaches for the real dataset, and the same procedure was then repeated using the simulated lesion dataset (i.e., total of 400 spatial normalizations). As an additional quality-assurance check, the resulting deformations were also applied to the corresponding lesion masks to evaluate how each processing pipeline handled focal white matter lesions. For each normalization approach, inter-subject variability (across normalized T1-weighted images) was quantified using both mutual information (MI) and coefficient of variation (COV), and the corresponding normalized lesion volumes were evaluated using paired-sample t-tests.

Results

All four nonlinear warping methods outperformed conventional linear normalization, with SPM (CAT12) yielding the highest MI values, lowest COV values, and proportionately-scaled lesion volumes. Although lesion-filling improved spatial normalization accuracy for each of the methods tested, these effects were small compared to differences between normalization algorithms.

Conclusions

SPM (CAT12) warping, ideally combined with lesion-filling, is recommended for use in future MS brain imaging studies requiring spatial normalization.

Introduction

Multiple Sclerosis (MS) is a neurodegenerative disorder of the central nervous system characterized by focal lesions and atrophy in both white matter (WM) and gray matter (GM) regions [1]. Several quantitative MRI methods such as diffusion tensor imaging (DTI) [2], magnetization transfer imaging (MTI) [3], and myelin water imaging (MWI) [4] have been widely adopted to study MS and other WM disorders [5,6]. Such studies are often predicated on precise one-to-one spatial mappings between the brain images of different individuals. This is typically achieved by warping each participant's brain images to a common template, which then facilitates voxel-wise or region of interest (ROI)-based comparisons between individuals or groups [7]. However, among participants with MS, widespread brain pathologies (e.g., focal WM and GM lesions, distributed WM and GM atrophy, altered normal appearing white matter signals, etc.) are likely to affect the accuracy of automated spatial normalization methods.

Several spatial normalization algorithms and brain templates have been developed for studies of neurologically healthy individuals [8], but no previous studies have systematically compared normalization methods to evaluate their performance in the presence of MS-related pathologies. Generally, these approaches aim to minimize differences between each participant's data and a template image, such as Talairach and Tournoux or Montreal Neurological Institute (MNI) templates [9], using linear and/or nonlinear spatial transformations [10,11]. Linear transformations apply the same translation, rotation, and scaling parameters to all voxels within an image. Although they are robust to local pathologies, they do not accurately match individual brain structures, particularly in WM and other sub-cortical regions that are of particular interest in MS [12]. Conversely, nonlinear transformations apply different scaling parameters to each voxel. This allows more localized region-specific deformations, but the high-dimensional nature of these algorithms renders them prone to over-fitting during the template matching process, which can reduce or eliminate abnormal but potentially salient features in the images (e.g., erroneously increasing or decreasing sizes of focal brain lesions) [7,12,13].

To minimize these types of erroneous spatial deformations, brain lesions in clinical populations are often identified and either discounted (aka, ‘de-weighted’) during spatial normalization or lesion-filled (aka, ‘in-painted’) by intensity-correcting them based on signals from neighboring, normal-appearing tissue before spatial normalization [14]. Lesion-filling has been shown to improve anatomical correspondence, as well as WM and GM volume measurements among participants with MS [15,16].

Although studies comparing different image processing pipelines and spatial normalization methods have revealed performance differences in Alzheimer's disease, mild cognitive impairment, drug-resistant epilepsy, and stroke [7,17], no such comparisons have been reported for MS. Therefore, in the absence of established ‘best-practice’ guidelines for spatially normalizing brain imaging data in the presence of MS pathologies [7,12], we aimed to: 1) use MS brain imaging data to systematically evaluate the performance of five commonly-used spatial normalization approaches in four popular neuroimaging software packages (MRIStudio, FSL, ANTs and SPM) both before and after lesion-filling; 2) create a simulated lesion dataset to compare how each normalization approach was specifically affected by focal lesions rather than global differences in brain volumes and GM/WM signal intensities; and 3) determine a data processing pipeline that would enable the most reliable comparisons between individuals or groups in MS neuroimaging studies.

Section snippets

Data acquisition

The current study used de-identified MRI images from 20 randomly selected participants with MS [16 female; 4 male] and one randomly selected neurologically healthy participant enrolled in the ongoing Comorbidity, Cognition and Multiple Sclerosis (CCOMS) Study [18]. The study was approved by our institutional Research Ethics Board and written informed consent was obtained from all participants. All participants with MS (19 relapsing-remitting; 1 secondary-progressive) were diagnosed by

Results

In total, 400 spatial normalizations were performed on the MRI datasets (20 images × 5 spatial normalization algorithms × 2 lesion-filling conditions × 2 patient/simulated datasets). Since some spatial normalization algorithms were more computationally intensive than others, ranging from ~1 min/participant for linear (affine) to ~60 min/participant for MRIStudio (LDDMM), approximate processing times for each approach and exact version numbers for each software package are shown in Table 1.

Discussion

We systematically compared five of the most commonly used and freely available spatial normalization methods, before and after lesion-filling in the presence of MS pathologies and found that nonlinear warping methods systematically outperformed conventional linear normalization in terms of their ability to consistently warp MS participant images to a common template. In particular, SPM12 (CAT12), closely followed by ANTs (SyN), yielded the most consistent results, with the highest MI and lowest

Conclusions

Our findings indicate that nonlinear warping methods systematically outperformed linear (affine) spatial normalizations and that the SPM (CAT12) algorithm proved to be the most robust among the methods tested, particularly when using lesion-filled images. We therefore suggest that future MS brain imaging studies use SPM (CAT12) and lesion-filling for optimal spatial normalizations. This should yield more accurate voxel-wise and ROI-based comparisons between MS individuals or groups within each

CRediT authorship contribution statement

Salina Pirzada: Conceptualization, Methodology, Investigation, Formal analysis, Data curation, Visualization, Writing - original draft, Writing - review & editing. Md Nasir Uddin: Conceptualization, Methodology, Software, Investigation, Formal analysis, Data curation, Visualization, Writing - original draft, Writing - review & editing. Teresa D. Figley: Investigation, Data curation, Visualization, Writing - review & editing. Jennifer Kornelsen: Writing - review & editing, Funding acquisition.

Acknowledgments

This work was supported by the Waugh Family Foundation Multiple Sclerosis Society of Canada operating grant (EGID-2639), The Canadian Institutes of Health Research (THC-135234), The Winnipeg Health Sciences Centre Foundation (HSCF), The Natural Sciences and Engineering Research Council of Canada (RGPIN-2016-05954), and the Brain Canada Foundation. RAM also receives support through a Manitoba Research Chair, as well as a Waugh Family Multiple Sclerosis Research Chair.

References (45)

1

Indicates equal contributions to the manuscript (joint first-authors).

2

The Comorbidity, Cognition and Multiple Sclerosis (CCOMS) Study Group includes: Ruth Ann Marrie, MD, PhD (University of Manitoba, Principal Investigator); John D. Fisk, PhD (Dalhousie University, Co-Principal Investigator); James Bolton, MD (University of Manitoba, Co-Investigator); Chase R. Figley, PhD (University of Manitoba, Co-Investigator); Lesley Graff, PhD (University of Manitoba, Co-Investigator); Jennifer Kornelsen, PhD (University of Manitoba, Co-Investigator); James J. Marriott, MD (University of Manitoba, Co-Investigator); Erin Mazerolle, PhD (University of Calgary, Co-Investigator);Ronak Patel, PhD (University of Manitoba, Co-Investigator); Teresa D. Figley, MSc (University of Manitoba, Research Coordinator); Carl A. Helmick, BCS (Dalhousie University, MRI Data Analyst); Md Nasir Uddin, PhD (University of Manitoba, MRI Physicist and Data Analyst); Charles N. Bernstein, MD (University of Manitoba, Collaborator).

View full text